Knowledge Retrieval

Knowledge Retrieval

This article lists the capabilities of the Jan platform and guides you through using RAG to chat with PDF documents.


To access this feature, please enable Experimental mode in the Advanced Settings.

Enable the Knowledge Retrieval

To chat with PDFs using RAG in Jan, follow these steps:

  1. Create a new thread.
  2. Click the Tools dropdown menu.
  3. Ensure that Retrieval is toggled on.
  4. Adjust the Retrieval settings as needed. These settings include the following:
Retrieval- Utilizes information from uploaded files, automatically retrieving content relevant to your queries for enhanced interaction.
- Use this for complex inquiries where context from uploaded documents significantly enhances response quality.
Embedding Model- Converts text into numerical representations for machine understanding.
- Choose a model based on your needs and available resources, balancing accuracy and computational efficiency.
Vector Database- Facilitates quick searches through stored numerical text representations to find relevant information efficiently.
- Optimize your vector database settings to ensure quick retrieval without sacrificing accuracy, particularly in applications with large data sets.
Top K- Determines the number of top-ranked documents to retrieve, allowing control over search result relevance.
- Adjust this setting based on the precision needed. A lower value for more precise, focused searches and a higher value for broader, more comprehensive searches.
Chunk Size- Sets the maximum number of tokens per data chunk, which is crucial for managing processing load and maintaining performance.
- Increase the chunk size for processing large blocks of text efficiently, or decrease it when dealing with smaller, more manageable texts to optimize memory usage.
Chunk Overlap- Specifies the overlap in tokens between adjacent chunks to ensure continuous context in split text segments.
- Adjust the overlap to ensure smooth transitions in text analysis, with higher overlap for complex texts where context is critical.
Retrieval Template- Defines the query structure using variables like {CONTEXT} and {QUESTION} to tailor searches to specific needs.
- Customize templates to closely align with your data's structure and the queries' nature, ensuring that retrievals are as relevant as possible.
  1. Select the model you want to use.

To upload an image or GIF, ensure that you are using a multimodal model. If not, you are limited to uploading documents only.

  1. Click on the 📎 icon in the chat input field.
  2. Select Document to upload a document file.